Huang: ship AI like an omelet—hold it back until rigorously tested; OpenAI's Astra got 2x more on Rubin

sudoraohacker · x · 2026-09-18

Fireside Alpha compiled key quotes from OpenAI Safety Week interviews:

Jensen Huang: shipping AI follows the same rule as shipping an omelet or an airplane engine—go as fast as you can, but no faster than readiness. Most incidents trace to premature releases; contained testing labs and rigorous engineering are the answer.
Sachin Katti: OpenAI ran Blackwell-optimized code for Astra on NVIDIA's Rubin chip and got 3x throughput out of the box; then let Astra act as an agent optimizing its own inference on Rubin, yielding another 2x in 72 hours—work that previously took weeks or months.
Safety costs compute: alignment and interpretability research require substantial extra compute. OpenAI disclosed that after strengthening chain-of-thought monitoring following the Hugging Face incidents, monitoring alone consumes 20% as much compute as the underlying model.

More from Companies & People